You read something at eleven. By three you know it existed and nothing else: not the app, not the sender, not the words. Cmd+F only searches the window in front of you. Screen memory is the fix for that specific failure: a record of the text your computer showed you, searchable later, with a way back to the moment.
Update, 26 October 2026: Sharpy, the app this post describes, is now an AI to-do list for Mac built on this kind of memory. What it reads and where the text lives are the same; what it does with it is below, updated.
What gets captured
There are two ways to build it.
Frames. Record the screen as images or video, run OCR, index the words. This is what Rewind did and what Screenpipe does. You get pictures back, at the cost of disk and the fact that pictures of your screen exist on the machine.
Text. Ask macOS for the text of the front window through the Accessibility API, the same channel screen readers use, and store the words. This is what Sharpy does, several times a minute, for the app in front. No video, no screenshots, no keystrokes. What it keeps looks like a page of notes with a timestamp, app name, window title and URL, not a photo.
Text capture has one real limit: apps built on Electron, which includes many chat and AI desktop clients, expose their text less completely than native apps, so a passage can come back partial. Web pages in a browser usually read better than the same product’s desktop app.
What stays on the Mac
In Sharpy, everything captured stays in ~/Library/Application Support/Sharpy/ on your Mac. Erase everything in Settings, and it is gone.
Before a line is written, card numbers, keys, phone numbers and codes are masked. Passwords, banking, logins and private windows are never saved at all. You can pause capture, leave out any app, site or window, and open the Privacy page to see what was kept and remove it.
What a model sees
Storing is local. Understanding is not. When you double-tap Option and ask “what did we agree with Alex last time?”, Sharpy searches the local history, opens the matching moments, and sends those passages, not the whole history, to a cloud model through Sharpy’s gateway. The answer comes back with citation chips; click one and the exact moment opens, so you can check the claim against the source.
The same applies to the to-do list Sharpy writes from that text and to the writing feature (tap the right Option in any text field, get a draft in your tone that knows the thread). Relevant context travels; providers may keep it up to 30 days for abuse monitoring.
What you can do with it once it exists
- A list that writes itself. “I’ll send the deck tomorrow” in Slack becomes a task due tomorrow, linked to the message. A sent reply or a merged pull request crosses it off.
- Find. “Where did I see the invoice number for the Astana contract?” The answer opens the message.
- Resume. “What was I doing before lunch?” The last few sessions, with the draft still open.
- Write. Tap the right Option in a reply field. The draft knows what the thread decided and sounds like you.
- Share with your agent. Claude Code, Codex and Cursor read the same list and memory over MCP; see Claude Code memory: what it keeps, and what it can’t see.
Who it is for
People whose work is spread across chats, documents and a browser, who explain the same context for the fifth time each week. Founders. Developers living in Claude Code, Telegram and Slack. Anyone who used Rewind until it stopped in December 2025.
Sharpy runs on macOS 14 or later on Macs with Apple silicon, one Mac per account, and is free in early access. Windows has a waitlist. Get early access.
